Classifier Selection Based on the Correlation of Diversity Measures: When Fewer Is More

The ever-growing access to high-resolution images has prompted the development of region-based classification methods for remote sensing images. However, in agricultural applications, the recognition of specific regions is still a challenge as there could be many different spectral patterns in a same studied area. In this context, depending on the features used, different learning methods can be used to create complementary classifiers. Many researchers have developed solutions based on the use of machine learning techniques to address these problems. Examples of successful initiatives are those dedicated to the development of learning techniques for data fusion or Multiple Classifier Systems (MCS). In MCS, diversity becomes an essential factor for their success. Different works have been using diversity measures to select appropriate high-performance classifiers, but the challenge of finding the optimal number of classifiers for a target task has not been properly addressed yet. In general, the proposed solutions rely on the a priori use of ad hoc strategies for selecting classifiers, followed by the evaluation of their effectiveness results during training. Searching by the optimal number of classifiers, however, makes the selection process more expensive. In this paper, we address this issue by proposing a novel strategy for selecting classifiers to be combined based on the correlation of different diversity measures. Diversity measures are used to rank pairs of classifiers and the agreement among ranked lists guides the classifier selection process. A fusion framework has been used in our experiments targeted to the classification of coffee crops in remote sensing images. Experiment results demonstrate that the novel strategy is able to yield comparable effectiveness results when contrasted to several baselines, but using much fewer classifiers.